Clim Dyn DOI 10.1007/s00382-017-3594-0

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial shock effect Ravi P. Shukla1   · Bohua Huang1,2 · L. Marx1 · James L. Kinter1,2 · Chul‑Su Shin1,2 

Received: 3 August 2016 / Accepted: 18 February 2017 © Springer-Verlag Berlin Heidelberg 2017

Abstract  This study evaluates the seasonal predictability of the Indian summer monsoon (ISM) rainfall using the Climate Forecast System, version 2 (CFSv2), the current operational forecast model for subseasonal-to-seasonal predictions at the National Centers for Environmental Prediction (NCEP). From a 50-year CFSv2 simulation, 21 wet, dry and normal ISM cases are chosen for a set of seasonal “predictions” with initial states in each month from January to May to conduct predictability experiments. For each prediction, a five-member ensemble is generated with perturbed atmospheric initial states and all predictions are integrated to the end of September. Based on the measures of correlation and root mean square error, the prediction skill decreases with lead month, with the initial states with the shortest lead (May initial states) generally showing the highest skill for predicting the summer mean (June to September; JJAS) rainfall, zonal wind at 850  hPa and sea surface temperature over the ISM region in the perfect model scenario. These predictability experiments are used to understand the finding reported by some recent studies that the NCEP CFSv2 seasonal retrospective forecasts generally have higher skill in predicting the ISM rainfall anomalies from February initial states than from May ones. Electronic supplementary material  The online version of this article (doi:10.1007/s00382-017-3594-0) contains supplementary material, which is available to authorized users. * Ravi P. Shukla [email protected] 1

Center for Ocean‑Land‑Atmosphere Studies (COLA), George Mason University, 270 Research Hall, Mail Stop 6C5, 4400 University Drive, Fairfax, VA 22030, USA

2

Department of Atmospheric, Oceanic, and Earth Sciences, George Mason University, Fairfax, VA, USA





Comparing the May climatologies generated by the February and May initialized CFSv2 retrospective forecasts, it is found that the latter shows larger bias over the Arabian Sea, with stronger monsoon winds, precipitation and surface latent heat flux. Although the atmospheric bias diminishes quickly after May, an accompanying cold bias persists in the Arabian Sea for several months. It is argued that a similar phenomenon does not occur in the predictability experiments in the perfect model scenario, because the initial shock is negligible in these experiments by design. Therefore, it is possible that the stronger model bias and initial shock in the May CFSv2 retrospective forecasts over the Arabian Sea may be a major factor in affecting ISM prediction skill. Keywords  Indian summer monsoon · NCEP CFSv2 · Implication of the initial shock effect · Predictability

1 Introduction Most of the Indian subcontinent receives 60–90% (Joshi and Rajeevan 2006) of the annual rainfall during the Indian summer monsoon (ISM) season (June to September, JJAS). The interannual variability (IAV; Shukla 1987) of ISM rainfall produces strong monsoon rainfall seasons that may lead to floods and weak monsoon rainfall seasons that is associated with droughts. Even small fluctuations in the seasonal JJAS rainfall strength and spatial distribution can have devastating impacts on India’s agricultural sector, water resource management and economy (Parthasarathy et al. 1988; Webster et al. 1998; Gadgil and Gadgil 2006). Therefore, early and reliable prediction of the IAV of ISM rainfall (ISMR) is highly desirable, especially to India’s policy makers. For more than a century, the prediction of

13

Vol.:(0123456789)



the IAV of ISMR has been based on statistical techniques (Walker 1923; Rajeevan et  al. 2000, 2004; Shukla et  al. 2011). On the other hand, Charney and Shukla (1981) provided a scientific basis for monsoon prediction beyond the limit of weather predictability and demonstrated that the interannual variability of tropical rainfall and circulation is influenced by slowly varying lower-boundary forcing such as sea surface temperature (SST), soil moisture, sea ice and snow cover. Wang et  al. (2005) showed that coupled ocean–atmosphere models are more appropriate to simulate Asian-Pacific summer monsoon rainfall accurately than atmospheric models with prescribed SST anomalies, because the latter are unable to reproduce the moderation of the SST fields through air-sea interaction that is important to the precipitation over the warm water in the Indo-Pacific region. Using hindcast experiments, Zhu and Shukla (2013) demonstrated that, in the absence of air-sea feedback, the atmospheric model produces higher rainfall biases and unrealistic interannual variability, suggesting that a coupled atmosphere–ocean prediction system is necessary for prediction of rainfall over Asia-Pacific region. It has been well established that interannual variability of the ISM system is associated with SST anomalies in the tropical Pacific, such as those associated with El Niño and the Southern Oscillation (ENSO; Rasmusson and Carpenter 1983; Shukla and Paolin 1983; Webster and Yang 1992; Shukla et al. 2011; Shukla and Kinter 2014; Shukla and Huang 2015b); in the Indian Ocean, such as the Indian Ocean dipole (IOD) (Saji et  al. 1999; Huang and Kinter 2002; Huang and Shukla 2007a, b; Wu and Kirtman 2004, 2005); and in the Arabian Sea (Saha 1970; Shukla 1975; Izumo et al. 2008; Shukla and Huang 2015a). Izumo et al. (2008) found that a decrease in Somalia–Oman upwelling strengthens monsoon rainfall along the western coast of India by increasing the SST along the Somalia–Oman coasts, and thus increasing moisture transport toward the Indian Western Ghats. Furthermore, Shukla and Huang (2015a) revealed that pre-monsoon conditions in the Arabian Sea may have a significant impact on the ISMR during early to mid-summer, which may be independent of the ENSO effect. Other studies also showed that the cold SST biases in the Arabian Sea and Indian Ocean in many coupled models significantly reduce the ISM precipitation and circulation (Levine and Turner 2012; Levine et al. 2013). Several studies (Achuthavarier et  al. 2012; Chaudhari et  al. 2013a; Shukla and Kinter 2014; Jiang et  al. 2013; Saha et  al. 2014a; Shukla and Huang 2015b) have discussed the challenges in simulating and predicting the seasonal mean ISM rainfall and its variability in both Climate Forecast System, version 2 (CFSv2), and its predecessor, CFSv1. Although CFSv1 does not reproduce the observed relationships between NINO3.4 SST index and Indian summer monsoon indices during and before the

13

R. P. Shukla et al.

summer monsoon season (Shukla and Kinter 2014), the CFSv2 is able to reproduce the correlation and do it more realistically to some extent (Shukla and Huang 2015b). On the other hand, CFSv2 still produces weaker monsoon precipitation over the Indian subcontinent, which may be connected with an excessive surface convergence over the central and eastern equatorial Indian Ocean that reduces the moisture transport toward the Indian subcontinent (Shukla and Huang 2015b). Jiang et  al. (2013) demonstrated that CFSv2 exhibits improvements in predicting the large-scale circulation and precipitation patterns of the Asian summer monsoon (ASM) due to the reduction of the cold bias over the Asian continent compared to CFSv1. Saha et  al. (2014a) showed that spatial patterns of seasonal mean rainfall and circulation over an extended Indian domain are more realistic in CFSv2 than in CFSv1. Previous studies have evaluated the predictive skill of ISMR by the NCEP CFSv2 hindcasts. In particular, Chattopadhyay et  al. (2015) and Saha et  al. (2016) found that retrospective forecasts made with NCEP CFSv2 from February (hereafter, Feb) initial conditions (IC) have higher ISMR prediction skill than those from May IC. To explain this result, Chattopadhyay et al. (2015) proposed that, over the equatorial central Pacific, CFSv2 has a very strong cold SST and dry rainfall bias, which is larger in the May IC runs than in the Feb IC runs, and that these biases are due to strong unrealistic coupled feedback in the central Pacific region. So, the ENSO-monsoon teleconnection pattern shifts westward with shorter lead-time, producing unrealistic patterns in comparison to observations, and causing a lower prediction skill with May IC. Saha et  al. (2016) found that retrospective forecasts at longer lead-time (lead3) have higher ISMR skill than at shorter lead-time (lead0). The authors (Chattopadhyay et  al. 2015; Saha et  al. 2016) argue that there is a physical mechanism for higher prediction skill with Feb ICs in CFSv2. It should be noted that prediction skill is distinct from potential predictability, that, on average, potential predictability decreases with lead time (Lorenz 1963; Shukla 1981; among many others), and that any departure from a monotonic decrease of prediction skill with lead time, as reported by Chattopadhyay et al. (2015) and Saha et al. (2016), requires an explanation. We have conducted potential predictability experiment (see Sect. 2 for detail of the experimental design) using CFSv2 and explored an alternative explanation for higher prediction skill with longer lead. Our results demonstrate that, in a perfect model scenario, the prediction skill decreases with lead month, with the initial states with the shortest lead (May IC) generally having the highest skill for predicting the JJAS rainfall, zonal wind at 850  hPa (U850) and SST over the ISM region. It will be shown that May-initialized NCEP CFSv2 retrospective forecasts have stronger monsoon winds, precipitation and surface latent heat flux in

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial…

May over the Arabian Sea and cooling near Somalia–Oman upwelling region in June in comparison to February-initialized forecasts. Also, the atmospheric bias diminishes quickly after May but an accompanying cold bias persists in the Arabian Sea for few months. It is important to note that, unlike in predictions initialized from observations, the initial shock is negligible in the perfect-model predictability experiments. Therefore, the impact from the initial shock in the May hindcasts may be a major factor in affecting ISM prediction skill. Many studies have shown that climate drift is a common problem in coupled models, which is defined as departure of model climatology or regime behavior from the observed (Sausen et al. 1988; Rahmstorf 1995; Slingo and Palmer 2011). There are different processes responsible for climate drift. One factor may be an imbalance of the energy fluxes that are exchanged during a model simulation. To avoid the influence of initialization shock, it is necessary to confirm during seasonal forecasts that stateof-the art coupled general circulation model components (i.e., atmospheric-model, ocean-model, land-model and others) are consistent with one another at the initial time of the seasonal forecasts (for example, reforecasts initialized in February or May). Without some procedure to suppress initial shock, the model will be dominated by the initialization shock for the first few months. There are several possible causes of initialization shock. One is the imbalance in the vertical fluxes of sensible or latent heat when any two components of the coupled model are initialized separately (Mulholland et al. 2015). Zhang (2011a, b) explored the impacts of minimized initial coupling shocks in coupled data assimilation on climate prediction over seasonalto-interannual to multi-decadal time scales using a simple pycnocline prediction model. Results of these experiments revealed that the coupled model initialization, in which all coupled model components are coherently adjusted by observations minimizes the initial coupling shocks, reduces the forecast errors on seasonal-to-interannual time scales and enhances the predictability of the model on all time scales. Mulholland et  al. (2015) discussed the origin and impact of initialization shocks in coupled model forecast on medium-range and seasonal forecasting timescales, they found that the impact from oceanic shocks has the potential to influence the seasonal time scale. In this paper, we evaluate the prediction skill of ISM by CFSv2 using Feb IC and May IC using a perfect model scenario and explore the impact of initial shock with Feb IC and May IC. First, we have conducted a long (50-year) free simulation using CFSv2 and performed a predictability experiment based on that run. Our analysis shows that retrospective forecasts in the perfect model scenario from May IC have higher skill than Feb IC runs. Second, we provide evidence that the impacts from initial shock may

be responsible for higher skill from Feb IC in the NCEP CFSv2 retrospective forecasts as reported by Chattopadhyay et al. (2015) and Saha et al. (2016). The rest of this paper is organized as follows. Section 2 describes the model details, experimental design, and observational datasets used in this study. In Sect.  3, we discuss the climatological features of ISM and the prediction skill of the CFSv2 in perfect-model experiments. The impact of initial shock is explored in both NCEP CFSv2 retrospective forecasts and predictability experiments in the Sect. 4. The conclusions are given in Sect. 5.

2 Model descriptions, experimental design and data sets The coupled model used in this study is CFSv2 (Saha et al. 2014b), which is composed of interacting atmospheric, oceanic, sea ice and land-surface component models. The atmospheric model is a version of the NCEP Global Forecast System at T126 horizontal resolution (105-km grid spacing) and 64 vertical levels in a hybrid sigma-pressure coordinate. The geophysical fluid dynamics laboratory (GFDL) modular ocean model (MOM) version 4.0 is the oceanic component of CFSv2, which is configured for the global ocean with a horizontal grid of 0.5° × 0.5° poleward of 30°S and 30°N and a meridional resolution increasing gradually to 0.25° between 10°S and 10°N. The vertical coordinate is geo-potential (z-) with 40 levels. The maximum depth is approximately 4.5  km. The oceanic and atmospheric components exchange surface momentum, heat and freshwater fluxes, as well as SST, every 30 min. We employed a slightly updated version of NCEP CFSv2 (see Huang et al. 2015 for details) for predictability experiments. Huang et al. (2015) eliminated a code inconsistency at the air-sea interface, which leads to a significant improvement especially during boreal summer in the model simulation. We have conducted a 50-year simulation using the CFSv2, which is initiated from the observed initial conditions in January 1980. Using the control simulation of the 50  years as a “nature run”, we have chosen 21 cases, which included typical wet, dry and normal monsoon and ENSO years, and conducted a set of predictability experiments. Each of these “predictability runs” starts with initial conditions at 00Z on the first day in each of the months from January to May taken from the 50-year simulation and is integrated to the end of September. Because the initial conditions are drawn from the nature run, the evolution of model fields in the predictability runs can be directly compared to the same quantities in the nature run. This is sometimes referred to as a “perfect model” scenario, because the climatology of the forecasts is identical to the climatology of the free simulation. To generate an ensemble of

13



five members, we perturbed the atmospheric fields at the initial time using a weighted average with the corresponding fields at the 00Z of the last day of the preceding month, while the ocean initial states remained the same as in the original simulation. In this paper, we will discuss results mainly related to Feb IC runs and May IC runs. In addition, we have also employed the 9-month retrospective hindcasts from the NCEP Climate Forecast System Reanalysis and Reforecast (CFSRR) Project (Saha et  al. 2014b; http://cfs.ncep.noaa.gov), which includes predictions initialized in all calendar months from January 1982 to December 2009. The retrospective forecasts have initial conditions (ICs) at 0000, 0600, 1200, and 1800 UTC on every 5th day, starting from 1 January of every year. Oceanic and atmospheric initial conditions are from the NCEP Climate Forecast System Reanalysis (CFSR; Saha et al. 2010). For this analysis, the ensembles mean predictions are comprised of 24 forecasts. As an example, for an ensemble monthly prediction with start month of February, the 24 ensemble members are the predictions from ICs on 11, 16, 21, 26, 31 January, and 5 February and each date has initial conditions at 0000, 0600, 1200, and 1800 UTC. The observed precipitation used in this study is the monthly Global Precipitation Climatology Project analysis (GPCP v2.2; Adler et al. 2003) at a resolution of 2.5° latitude by 2.5° longitude for the period 1979–2015. Monthly gridded data of U850 is from the NCEP/National Center for Atmospheric Research (NCEP/NCAR) reanalysis (Kalnay et  al. 1996). The National Oceanic and Atmospheric Administration (NOAA) extended reconstructed sea surface temperature (ERSST, version 4.0, 2° × 2° grid; Smith et al. 2008) has been used for the period 1965–2015. Monthly anomalies are derived for each dataset with respect to its own monthly climatology. Seasonal mean anomalies are then calculated from the monthly anomalies. The statistical significance of spatial prediction skill is measured by correlation coefficient and given the sample size, the correlation values at 90, 95, 98 and 99% confidence levels are 0.36, 0.42, 0.49 and 0.54, respectively.

3 Results 3.1 The spatial and temporal distribution of rainfall, U850 and SST Before discussing the seasonal prediction skill in the predictability runs, it is necessary to quantify CFSv2 fidelity in simulating the important climatological characteristics of seasonal (JJAS) mean precipitation, SST and low-level zonal wind (U850) over the ISM region in the 50-year free run. Figure  1 shows the 50-year mean JJAS precipitation (Fig. 1a), SST (Fig. 1c) and U850 (Fig. 1e) in the CFSv2

13

R. P. Shukla et al.

50-year free simulation in the ISM region and corresponding fields from the observations (Fig. 1b, d, f, respectively), averaged over the available record for each. The CFSv2 (Fig.  1a) captures the main observed precipitation pattern (Fig. 1b) over India and the surrounding oceans, including centers of action in the Western Ghats, the northern Bay of Bengal, and the equatorial Indian Ocean off the Sumatra coast. The region of large rainfall over the oceanic warm pool encompassed by the 28 °C isotherm of SST is noted in both CFSv2 and observations (Fig. 1c, d, respectively). The magnitude and structure of observed JJAS SST (Fig.  1d) over the Arabian Sea, Bay of Bengal and Indian Ocean is remarkably well captured by CFSv2 (Fig.  1c). Consistent with that, U850 is also well simulated in the model (Fig. 1e), especially over the central Arabian Sea, the Indian subcontinent region and the Bay of Bengal, as well as the easterly wind over the Indian Ocean between 22°S and 5°S. The westerly wind over the Somalia coast is approximately 9–14 m/s in the NCEP/NCAR reanalysis, while the corresponding values in the CFSv2 simulation is approximately 9–12  m/s. The simultaneous correlation coefficient (CC) between the normalized time series of extended Indian summer rainfall index (EISMRI; averaged over the area 8°N–25°N, 70°E–85°E, which is the box shown in Fig. 1a) (Fig.  2a, b; black curve) and JJAS NINO3.4 SST index (Fig.  2a, b; red curve) in CFSv2 is −0.58, which may be compared to −0.40 in the observations. We may conclude that EISMRI in both the CFSv2 simulation and observations significantly depends on ENSO. Although the CFSv2 simulation captures all the important features of JJAS precipitation, SST and U850 over the ISM region, it underestimates (overestimates) precipitation over the Indian landmass (Indian Ocean region) (Fig.  1g) (Shukla and Huang 2015b; Pokhrel et  al. 2016). We have found that the rainfall is considerably overestimated in the northeast part of India and the over the Arabian Sea (Fig. 1g). The low-level winds are also weak across the southwest of India (Fig. 1i). The details of CFSv2 discrepancies in the ISM, viewed in a global context, are discussed in Shukla and Huang (2015b). Based on the normalized JJAS (June to September) extended Indian summer monsoon rainfall index (EISMRI; Fig.  2a; black line) for 50  years, we have selected seven strong monsoon years in which the rainfall is greater than +1 standard deviation, five weak monsoon years in which the rainfall is less than −1 standard deviation, and five normal monsoon years in which the rainfall is close to the mean. In addition, based on the normalized JJAS NINO3.4 SST index (Fig.  2a; red line), we have chosen five warm, five cold and five normal ENSO years. A few years are common between the two sets of 15, so, the total number of cases is 21. The numbers of strong, weak and normal monsoon years are seven, five and five respectively and there are four cases when normalized

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial…

(a)

(b)

(g)

(c)

(d)

(h)

(e)

(f)

(i)

Fig. 1  The spatial distributions of seasonal (JJAS) mean climatology of rainfall (colored shading) in a 50-year control simulation of CFSv2, b GPCP. The scale for the magnitude in “mm/day” is shown at upper right. c, d as in a, b but for JJAS SST. The scale for the magnitude in “degree Celsius” is shown below these panels. e, f as in a,

b but for zonal wind at 850 hPa (U850). The scale for the magnitude in “m/s” is shown at lower right. The climatological biases relative to observations for rainfall (g), SST (h) and U850 (i). The scale for the biases for all cases is shown below these panels

EISMRI show values around ± 0.5. These cases are chosen for a set of seasonal “predictions” with initial states from January to May to conduct predictability experiments. For each prediction, a five-member ensemble is

generated with perturbed atmospheric initial states and all predictions are integrated to the end of September. The number of cases was limited to 21 due to computational constraints. Figure S1 (see supplemental material)

13



R. P. Shukla et al.

(a)

(b)

Fig.  2  a Year-to-year variation of standardized JJAS extended Indian summer monsoon rainfall index (EISMRI: black line; box in Fig. 1a: 8°N–25°N, 70°E–85°E) and JJAS NINO3.4 (red line) index in 50-year control simulation of CFSv2. We have indicated seven strong monsoon (SM; greater than +1 standard deviation), five weak mon-

soon (WM; less than −1 standard deviation) and five normal monsoon (NM; close to the mean) years. There are four cases when normalized EISMRI show values around ±0.5, indicated by “+”. Years on each panel in map are model years. b as in a but for observation

shows the ensemble mean spatially-averaged U850 over the northern Indian Ocean (U850 index or U850I; box in Fig.  1e: 0°–15°N, 50°E–90°E) for both Feb IC and May IC runs and nature run values. Remarkably, both February and May initialized predictability runs in this perfect model scenario capture the development and transitions of U850I during the ISM season in all 21 cases. However, the magnitudes of U850I anomalies (Fig.  3) are clearly underestimated in both sets of predictability runs during model years 1985, 1992, 2004, 2024, 2025 and 2028 (note that the model years are numbered beginning with 1980, because the nature run was initialized from observations in 1980; however, the results in a given year of the nature run or any of the predictability runs are not comparable to the observations), although the May IC runs seem to capture the magnitude slightly better than the Feb initialization during 1985, 1992 and 2004. In order to quantify the CFSv2 fidelity in predicting JJAS rainfall over the ISM region with both Feb and May IC, we examine area-averaged rainfall for the ensemble

means of both sets of predictability runs and the nature run over the extended Indian region (EISMRI: box in Fig.  1a: 8°N–25°N, 70°E–85°E) as displayed in Fig. S2 (see supplemental material). Both sets of predictability runs capture the onset, development and withdrawal of the monsoon rainfall in strong, weak and normal monsoon years (Fig. S2). However, there are clear distinctions between predictions with both predictability runs and the nature run (Fig. S2), mainly in the 1983, 1999, 2005, 2015, 2016, 2025 and 2028 model years. This may imply that, even in the perfect model scenario and using perfect ocean initial conditions, the mean states of ISM rainfall is not easy to predict. In order to get further insight into the predicted EISMRI, we examined EISMRI anomalies in both predictability runs (Fig. 4). It is also clear that some anomalous episodes were missed by both sets of predictability runs. For instance, both Feb and May IC predictability runs produced unrealistic anomalies in model years 1983, 2004 and 2015, while the magnitude of the strong monsoon in 2005 was missed by both Feb and May initializations (Fig. 4).

13

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial…

Fig. 3  Time series of monthly anomalies U850 (black line; February to September) over the northern Indian Ocean (box in Fig.  1e; domain: 0°–15°N, 50°E–90°E) for 21 selected cases in the 50-year control simulation of CFSv2. Colored curves are forecast of U850 anomalies during the 21 cases initialized with February (red line)

and May (blue line) initial conditions (ICs). We have indicated seven strong monsoon (SM), five weak monsoon (WM) and five normal monsoon (NM) years. Years on each panel in map are model years. The unit of U850 is m/s

3.2 Prediction skill of JJAS rainfall, U850 and SST anomalies

calculated between February (or May) initialized reforecasts for 21 cases and corresponding 21 cases in 50-year control simulation during Indian summer monsoon season (JJAS). The ACC map (Fig.  5b) with May IC shows significant skill over the Bay of Bengal, central Arabian Sea and western coast of India, northern and eastern India. Prediction skill with the Feb IC is apparent over some parts of northern India (Fig. 5a). The predictions exhibit a clear difference between the ACC patterns using Feb ICs and May ICs over the Indian landmass and its adjacent oceanic regions. The model also displays significant skill in the equatorial regions over western Indonesia with both Feb IC and May IC, which may be related to SST variations in the eastern tropical Pacific. Consistent with this result, the model shows higher prediction skill for JJAS U850 anomalies over the tropical Indian Ocean (TIO) and the northern

Since the same forecast system is used to conduct both the nature run and the predictability runs, the predictability experiments are based on a perfect model scenario. Moreover, since both May ICs and Feb ICs are generated from relatively small atmospheric perturbations on the instantaneous states of the control run, the impact of initial shock (Zhang 2011a, b) is negligible in this case, as will be discussed in Sect. 4. The skill differences of monsoon prediction between Feb IC and May IC are measured by the anomaly correlation coefficient (ACC; Fig.  5) and root mean square error (RMSE; Fig.  6) of seasonal JJAS rainfall, U850 and SST over the ISM region. Anomaly correlation and root mean square errors (RMSEs) has been

13



R. P. Shukla et al.

Fig. 4  Time series of monthly anomalies extended Indian monsoon rainfall (EISMR; black line; February to September) index (box in Fig.  1a; domain: 8°N–25°N, 70°E–85°E) for 21 selected cases in the 50-year control simulation of CFSv2. Colored curves are forecast of EISMR index anomalies in 21 cases initialized with February

(red line) and May (blue line) ICs. We have indicated seven strong monsoon (SM), five weak monsoon (WM) and five normal monsoon (NM) years. Years on each panel in map are model years. The unit of rainfall is mm/day

Arabian Sea in May IC (Fig. 5d) in comparison to Feb IC (Fig. 5c). The Somali low-level jet transports moisture from the southern Indian Ocean to the core monsoon region (Findlater 1969) through the Arabian Sea and TIO. It is remarkable that the model successfully predicts zonal wind using May IC over the Arabian Sea and TIO better than in the Feb IC cases. There have been several studies (Saha 1970; Shukla 1975; Rao and Goswami 1988; Saji et  al. 1999; Izumo et al. 2008; Shukla and Huang 2015a) demonstrating that the SST anomalies over the Indian Ocean and Arabian Sea play an important role in influencing ISMR variability. As expected, the prediction skill is higher for JJAS SST anomalies using May IC (Fig. 5f) than using Feb IC (Fig.  5e) over the entire Arabian Sea, TIO and Bay of Bengal. Within the Indian Ocean basin, the model shows higher correlations in both Feb IC and May IC over the

southwestern TIO (above 0.7 in both cases), where ENSO plays a significant role as a remote forcing and ocean dynamics is necessary in determining the SST anomaly evolution (Huang and Kinter 2002; Xie et al. 2002). The RMSE pattern for JJAS rainfall from May IC (Fig. 6a) depicts similar features as previously noted for the mean state (Fig. 1a). For example, then RMSE of rainfall is generally larger over the Western Ghats of India, the northern Bay of Bengal, and the TIO. Figure 6b, d, f displays the difference in RMSE between Feb IC and May IC for JJAS rainfall, U850 and SST, respectively. The predictions have generally higher RMSE for summer rainfall (Fig. 6b), U850 (Fig. 6d), and SST (Fig. 6f) using Feb IC than for May IC over the entire ISM region. The RMSE of SST with Feb IC (Fig.  6f) is above 0.6 °C over the southwestern TIO. The ACC for JJAS SST is higher in the model using May IC

13

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial…

(Fig.  5f) than using Feb IC (Fig.  5e) at the Somali coast around ­3oS to 3­ oN but the RMSE is slightly higher in the May IC runs. This implies that the May IC runs slightly overestimate the magnitude of JJAS SST anomaly in this region. We have found that May IC depicts higher prediction skill for JJAS rainfall, SST, U850 over Indian summer monsoon region and lower RMSE in comparison to February IC in the perfect model scenario. Figure  7 shows the anomaly pattern correlations and RMSE for JJAS precipitation over the Indian landmass only. Correlations (Fig. 7a) are generally higher in the May IC predictability runs (blue line) than Feb IC (red line) w.r.t corresponding years in the nature run. Consistent with this result, the RMSE (Fig. 7b) is generally higher for the Feb IC runs compared to the May IC runs. A similar conclusion can be drawn for the extended Indian region (60°E–100°E, 5°N–25°N) in terms of anomaly pattern correlations and RMSE of summer monsoon rainfall (Fig. S3). Now, we will explore the correlation and RMSE (Table 1) of ISM rainfall over three regions. Index-1 (D1) is defined as the area-averaged rainfall over a slightly extended Indian domain (box in Fig.  1a; 70°E–85°E, 8°N–25°N); whereas Index-2 (D2) is the area-averaged rainfall over a larger extended Indian domain (60°E–100°E, 5°N–25°N). Finally, Index-3 (D3) is the area-averaged rainfall over the Indian landmass region. The correlation between predictability runs and the nature run over D1 for May ICs (Feb ICs) is 0.76 (0.60), while it is 0.81 (0.69) for D2 and 0.73 (0.54) for D3. Based on the correlation analysis, the interannual variability of rainfall over these three regions is a better match for May IC than for Feb IC. The RMSE is higher over these domains in Feb IC than May IC (Table 1). These results are robust and not sensitive to reasonable variations of domains. The prediction skill in these CFSv2 predictability experiments using Feb IC and May IC shows that there is higher skill over the ISM region at shorter lead times (May IC) when there is no influence from coupled model initial shock (Zhang 2011a, b). We have discussed prediction skill during strong and weak monsoon years in the perfect model scenario. Figure S4 (see supplemental material) depicts JJAS seasonal composites of strong and weak monsoon years based on the EIMRI for the control simulation (Fig S4a, b), February initialized predictability runs (Fig S4c, d), and May initialized predictability runs (Fig S4e, f). We have removed the 50-year JJAS climatological mean of the control simulation in all cases. For example, the composite anomaly of strong monsoon years (Fig. S4c) is the difference between the JJAS mean of the seven cases with a strong monsoon for the reforecasts initialized in February and the 50-year JJAS climatology. During a strong monsoon season, the JJAS mean composites from the control simulation (Fig. S4a) have positive anomalies over most of India and the

surrounding oceanic region. Predictability runs initialized in both February (Fig. S4c) and May (Fig. S4e) have positive anomalies over the monsoon region, but, on closer inspection, the composite of predictability runs initialized in May (Fig. S4e) shows better agreement with the control simulation composite (Fig. S4a). The spatial correlation between the control simulation composite and the strong monsoon composites initialized in February and May, for the slightly extended Indian domain (black box in Fig. S4; domain (D1): 8°N–25°N, 70°E–85°E), are 0.63 and 0.77, respectively. During weak monsoon years, the composite from the control simulation (Fig. S4b) has negative anomalies over the core monsoon region. The weak monsoon composite for predictability runs initialized in May (Fig. S4f) is more similar to the control simulation composite (Fig. S4b) than the same composite for predictability runs initialized in February (Fig. S4d). The spatial correlation between the predictions and the control simulation for weak monsoon composites over the D1 region for predictability runs initialized in February (May) is 0.65 (0.80). Based on these results, it seems that the prediction skill of predictability experiments initialized in May for both strong and weak monsoon composites are similar. The same is true for predictability experiments initialized in February. The spatial structure of composites for normal monsoon years is not coherent (not shown) so we have not discussed the prediction skill of normal monsoon years here.

4 Discussion The above results demonstrate that CFSv2 has higher skill using May IC than Feb IC in predictions in the perfect model scenario and perfect ocean initial state. Why, then, is there higher prediction skill in the NCEP CFSv2 hindcasts initialized with Feb IC than the hindcasts initialized with May IC, as reported in Chattopadhyay et  al. (2015) and Saha et  al. (2016)? One fundamental difference between the predictability runs and the hindcasts is that the latter are strongly influenced by the model climate drift, including a large initial shock (Zhang 2011a, b), while those influences are practically absent in the former. Therefore, we hypothesize that the stronger influence of the initial shock in the May CFSv2 hindcasts is the main reason for its reduced prediction skill. Specifically, we demonstrate the impacts of coupled model initial shock in surface latent heat flux (SLHF), SST, rainfall and 10  m-wind over the Arabian Sea, both qualitatively and quantitatively on the monthly time scales. In particular, the large initial imbalance in the model atmospheric states over the Arabian Sea before the monsoon onset seems to generate lasting effects, possibly through its imprint on SST. Several papers have documented that the

13



R. P. Shukla et al.

(a)

13

(b)

(c)

(d)

(e)

(f)

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial… ◂Fig. 5  Anomaly correlation coefficients of JJAS rainfall anomalies

for a February ICs and b May ICs for 21 years prediction in CFSv2 w.r.t 21 selected cases in the 50-year control simulation of CFSv2. c, d as in a, b but for JJAS U850 anomalies. e, f as in a, b but for JJAS SST anomalies. Correlation coefficients at 90, 95, 98 and 99% confidence levels are 0.36, 0.42, 0.49 and 0.54 respectively

Arabian Sea is a pathway for moisture transport from the Indian Ocean into major ISM region (Saha 1970; Shukla and Mishra 1977; Izumo et  al. 2008; Shukla and Huang 2015a, b). To demonstrate the potential impact of initial shock in NCEP CFSv2 hindcasts (Saha et  al. 2014b), Fig.  8 shows the climatological differences of SLHF (colored shading), rainfall (contours), and 10 m-wind (vectors) in May, June, July, and August between hindcasts initialized with Feb IC and May IC over the ISM region while the corresponding differences pattern for SST are shown in Fig.  9k–o for May to September for the period 1982–2009. Figure  9a–e displays the climatological bias of SST from May to September between hindcasts initialized with Feb IC and observed SST (Smith et  al. 2008) while corresponding patterns with May IC are exhibited in Fig.  9f–j. Qualitatively, it is found that SLHF is huge in May (Fig. 8a; colored shading) over the Arabian Sea in hindcasts initialized with May IC and 10 m-wind (Fig. 8a; vectors) is also stronger over this region. The model produces excessive rainfall (Fig. 8a; contours) in May over the Arabian Sea in CFSv2 hindcasts with May IC. The center of action in SLHF is located over the Arabian Sea while the center of action in precipitation is located slightly east of the SLHF center, that is, close to the Western Ghats of India. The amplitudes of SLHF (rainfall) at the major centers of action in May are approximately 20  W/m2 (3  mm/ day), larger in hindcasts initialized with May IC than Feb IC. This is consistent with the fact that, in May hindcasts, the 10 m-wind is stronger over the center of action in SLHF by 2  m/s. Due to excessive SLHF and stronger winds in May over the Arabian Sea, the CFSv2 hindcasts with May IC tend to produce lower SST in the Arabian Sea in June (Fig. 9g). The SST difference in May (Fig. 9k) between the hindcasts with Feb IC and May IC is not large over the Arabian Sea but a significant difference in SST occurs east of the Somalia–Oman upwelling region during June (Fig. 9l) with magnitude >0.7 °C. Colder SST bias near the Somalia–Oman upwelling region in June using May IC (Fig. 9g) is larger than for Feb IC (Fig. 9b). This colder bias reduces moisture transport towards the core Indian monsoon region (Izumo et al. 2008). The large perturbations in the atmosphere in May, including the excessive rainfall over the Arabian Sea in hindcasts with May IC, has reduced in June (Fig. 8b). During July and August (Fig. 9h, i), colder SST continues in the entire Arabian Sea in hindcasts initialized with May IC, although the initial shock in SLHF,

rainfall and 10 hPa-winds has largely diminished (Fig. 8c, d). By August, the sign of the difference of SLHF (rainfall) between Feb IC and May IC is changed over the Arabian Sea (western coast of India) (Fig.  8d). Due to the longer lead time, the hindcasts from Feb IC should have experienced more climate drift. However, we attribute the large systematic difference in May between the May and February initialized hindcasts to the larger model initial shock in the former because it has larger bias. On the other hand, the climatological differences between the Feb IC and May IC predictions are negligible in the predictability experiments during May to August over the Arabian Sea (not shown). This is consistent with the fact that, with a perfect model and small initial perturbation, the model initial shock is negligible. To quantify the impact of initial shock, the climatological mean SLHF from the hindcasts for Feb IC and May IC is area-averaged over the central Arabian Sea (8°N–16°N; 54°E–74°E) in both CFSv2 hindcasts and predictability runs. The difference is displayed in Fig.  10a, b while the corresponding rainfall over southern India (8°N–18°N; 60°E–80°E) is shown in Fig.  10c, d. The difference in SLHF in May is 17.3 W/m2 in the NCEP CFSv2 hindcasts (Fig.  10a) and 1.5  W/m2 for predictability runs initialized with May IC (Fig.  10b). Consistent with this, the rainfall difference is 2.10  mm/day (Fig.  10c) in CFSv2 hindcasts from May IC, compared with 0.19 mm/day in the predictability runs (Fig. 10d). The initial shock in SLHF (rainfall) is reduced by 8.15 W/m2 (1.24 mm/day) from May to June (Fig. 10a, c) in CFSv2 hindcasts with May IC. During July and August, the total area-averaged rainfall is higher in CFSv2 hindcasts with Feb IC than May IC (Fig. 10c). On the other hand, the model bias in rainfall (Fig. 10d) is negligible in the predictability experiments by design. Figure 11a shows the area-averaged climatological SST in the Somalia–Oman upwelling region (outlined by the box in Fig. 9) in observations (black line) and CFSv2 hindcasts with Feb IC (red line) and May IC (blue line). The tendency of SST over the Somalia–Oman upwelling region is reproduced by both sets of hindcasts as in the observations, but the magnitude of SST is underestimated in both Feb IC and May IC cases. Due to stronger model bias in May in SLHF and rainfall (Fig. 10a, c), this difference for SST is larger in CFSv2 hindcasts with May IC during June to August (Fig. 11a). On the other hand, there is practically no initial shock in the predictability runs, so the corresponding averaged SST over the Somalia–Oman upwelling region displays almost equal tendency and magnitude in the nature run (black line) and the predictability runs with Feb IC (red line) and May IC (blue line) (Fig. 11b). These results demonstrate that May initialized NCEP CFSv2 hindcasts have a larger bias over the Arabian Sea in May, with stronger monsoon winds, precipitation and surface

13



R. P. Shukla et al.

(a)

(c)

(d)

(e)

(f)

Fig.  6  a Root mean square error (RMSE) skill of JJAS rainfall using May ICs for 21 years prediction in CFSv2 w.r.t 21 selected cases in the 50-year control simulation of CFSv2. The scale for the magnitude is shown below these panels. b Difference in RMSE between Feb ICs

13

(b)

and May ICs for JJAS rainfall. The scale for the magnitude is shown at middle right. c, d as in a, b but for JJAS U850. e, f as in a, b but for JJAS SST. The units of rainfall, U850 and SST are mm/day, m/s and °C respectively

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial…

(a)

(b)

Fig.  7  a Anomaly pattern correlation and b RMSE of JJAS rainfall from February ICs (red line) and May ICs (blue line) over Indian landmass region only during 21 years prediction using CFSv2 w.r.t corresponding 21 selected cases in the 50-year control simulation of CFSv2 Table 1  Correlation and RMSE between predicted JJAS rainfall with February initial conditions and May initial conditions during 21 cases and corresponding years in 50-years control simulation over three domains which are defined as D1: area averaged rainfall over

slightly extended Indian domain (box in Fig. 1a; domain: 8°N–25°N, 70°E–85°E); D2: area averaged rainfall over extended Indian domain (60°E–100°E, 5°N–25°N); D3: area averaged rainfall value over Indian landmass region

Correlation coefficient

Feb IC May IC

Root mean square error (RMSE)

D1

D2

D3

D1

D2

D3

0.60 0.76

0.69 0.81

0.54 0.73

1.21 1.01

0.71 0.59

1.03 0.87

latent heat flux, which it can affect the prediction skill of ISM rainfall (DelSole and Shukla 2010). We have calculated the area-averaged rainfall over the Indian landmass during Indian summer monsoon (June to September; JJAS) for observations and NCEP CFSv2 reforecasts initialized in May for each year in 1982–2009. Figure S5 depicts the bias between May initialized NCEP CFSv2 reforecast and observation year-by-year. The small and large bias subsets are formed as follows. The small-bias subset contains 15  years whose bias is less than or equal to 1.8 mm/day, and the large-bias subset contains 13 years whose bias is greater than 1.8  mm/day. We reorganized

the small- and large-bias subsets and observational data accordingly. We calculated the anomaly correlation and root mean square error (RMSE) between observations and May initialized NCEP CFSv2 reforecast for both subsets. The temporal correlation between the small-bias subset and observations over the Indian landmass is 0.89 and 0.56 for the large-bias subset. The RMSE is lower in the smallbias subset over the Indian landmass in comparison to the large-bias subset. This analysis provides evidence that May initialized NCEP CFSv2 reforecasts have higher prediction skill for JJAS rainfall over the Indian landmass when the bias is small.

13



R. P. Shukla et al.

As noted by an anonymous reviewer, the ensemble size, methods of ensemble generation and initialization, and the dates of initial conditions are different between the predictability experiments described here and the NCEP CFSv2 reforecasts (Saha et al. 2014b). To test the impact of at least some of these other factors, we have compared the results of the predictability runs with a set of reforecasts generated separately by the Center for Ocean-Land-Atmosphere Studies (COLA) using a revised version of CFSv2 (Huang et al. 2015; Shukla and Huang 2015b), referred to herein as CFSv2_rev. The revised model corrects a coding error in

Fig. 9  Spatial distribution of the climatological biases in the monthly ▸ SST predicted from the NCEP CFSv2 reforecast system with respect to ERSST v4 during the period 1982–2009 for the months of May, June, July, August, September from the February initialization (left panel; Fig. 8a–e) and May (middle panel; Fig. 8f–i) initialization. The scale for the magnitude in “degree Celsius” is shown at upper right. The climatological difference of SST in May to September between forecast initialized with Feb ICs and May ICs (Feb IC–May IC) over the Arabian Sea in the NCEP CFSv2 retrospective forecast is shown in the right panel (Fig. 8k–o). The scale for the magnitude in “°C” is lower right

(a)

(b)

(c)

(d)

Fig. 8  Spatial distribution of the climatological difference of SLHF (colored shading), rainfall (contours), 10 m-wind (vectors) in a May, b June, c July, d August between forecast initialized with Feb ICs and May ICs (Feb IC–May IC) over ISM region in the NCEP CFSv2 retrospective forecast (Saha et al. 2014b) for period 1982–2009. The

13

intervals of contours lines for rainfall are −3, −2.5, −2.0, −1.5, −1.0, −0.5, 0, 0.5, 1.0, 1.5, 2.0 and 2.5 mm/day and black contour denotes “0” value. The units of SLHF, rainfall and 10 m-wind are W/m2, mm/ day and m/s respectively

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial…

(a)

(f)

(k)

(b)

(g)

(l)

(c)

(h)

(m)

(d)

(i)

(n)

(e)

(j)

(o)

13



R. P. Shukla et al.

the surface fluxes exchanges between the ocean and atmosphere components at high latitude and adjusts the specified sea ice albedo to a higher value that avoids the problem of total melting of summer Arctic sea ice within a decade. The re-forecasts were made using ocean initial conditions (OIC) from the National Centers for Environmental Prediction (NCEP) Climate Forecast System Reanalysis (CFSR) (Saha et al. 2010), the NCEP Global Ocean Data Assimilation System (GODAS) (Behringer 2005), and the ECMWF Ocean Reanalysis System 3 (ORA-S3) (Balmaseda et  al. 2008), all for the period 1979–2008. For each OIC, four ensemble members were generated by using atmospheric and land surface initial conditions (AIC and LIC) taken from the instantaneous fields at 0000 UTC of the first 4 days in February and May of each year. The total number of ensemble members was 12−3 OICs  ×  4 AIC/LICs— for both February and May cases and all predictions were integrated to the end of September. We have calculated the prediction skill for precipitation in ISMRI (Indian landmass only) for both February initial states and May initial states for ensemble means of 4 members, 8 members and 12 members. The anomaly correlation between observations and the ensemble means of 4, 8 and 12 members initialized with February ICs are 0.44, 0.49, 0.49 respectively and 0.34, 0.44, 0.39 for initialized with May ICs. Therefore, the reforecasts generated in this way produce a similar result to that of Chattopadhyay et  al. (2015) and Saha

10

(a)

5

5

0

0

SLHF (W/m2)

2

SLHF (W/m )

10

et al. (2016), namely, that reforecasts using February initial states have higher prediction skill for ISMRI (Indian landmass only) than those using from May initial states. Using the CFSv2_rev hindcasts, we have also calculated the climatological differences of surface latent heat flux (SLHF) (colored shading), rainfall (contours), 10 m-wind (vectors) in May, June, July, August between forecasts initialized with Feb ICs and May ICs over the ISM region for ensemble means of 4 (Fig. S6a–d; see supplemental material), 8 (Fig. S6e–h) and 12 (Fig. S6i–l) members. The corresponding difference patterns of SST are shown for ensemble means of 4 members (Fig. S7 a–e), 8 members (Fig. S7 f–j) and 12 members (Fig. S7 k–o) during May to September for 1979–2008. We have found that results from CFSv2_ rev hindcasts for all cases (ensemble mean of 4, 8 and 12 members) are similar as discussed in Sect. 4 of this paper. For example, the SLHF losses are huge over the Arabian Sea, 10  m-wind is stronger over this region and there is excessive rainfall over the Arabian Sea in CFSv2_rev forecasts initialized with May IC for ensemble means of 4 (Fig. S6a), 8 (Fig. S6e) and 12 (Fig. S6i) members. The SST differences in May between the CFSv2_rev hindcasts with Feb ICs and May ICs are not large over the Arabian Sea, but there is a significant difference in SST (>0.7 °C) east of the Somalia–Oman upwelling region during June in all cases (Fig. S7b, Fig. S7g and Fig. S7l). There are several things in common between the “predictability experiments”

−5 −10 −15

(b)

−5 −10 −15

−20

May

June

July

−20

August

May

June

Months 1 0.5

1

(c)

0.5

−1 −1.5

July

August

(d)

−0.5 −1 −1.5

−2

−2

−2.5

−2.5 May

June

July

August

Months

Fig.  10  a The area average of climatological difference of SLHF (unit: W/m2) between forecast initialized with Feb ICs and May ICs in NCEP CFSv2 retrospective forecast and b for predictability experi-

13

August

0 rainfall (mm/day)

rainfall (mm/day)

0 −0.5

−3

July Months

−3

May

June Months

ments over the central Arabian Sea (8°N–16°N; 54°E–74°E) in May to August. c, d as in a, b but for rainfall (unit: mm/day) over southern extended Indian region (8°N–18°N; 60°E–80°E)

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial… Fig.  11  a The area averaged climatological SST near Somalia–Oman upwelling region (box in Fig. 9) in observation (black line) and forecast initialized with Feb ICs (blue line) and May ICs (red line) in NCEP CFSv2 retrospective forecast. b as in a but for predictability experiments. The unit of SST is °C

(a)

(b)

and the “CFSv2_rev hindcasts”. (1) Both experiments employed the same version of the CFSv2 model. (2) Both experiments used the same high-performance computing environment on the Extreme Science and Engineering Discovery Environment (XSEDE) platform. (3) Similar procedures were used to generate ensemble members in both experiments. (4) The initialization dates are identical in both experiments. (5) We have verified our conclusions using CFSv2_rev hindcasts for ensemble means of 4, 8 and 12 members. (6) No procedures to minimize the initial shock were employed in the CFSv2_rev hindcasts, and there were no documented procedures employed to

minimize the initial shock in the original CFSv2 hindcasts (Saha et al. 2014b). We have compared the prediction skill of ISMRI (Indian landmass only) in both predictability experiments and the CFSv2_rev hindcasts cases. We have found that prediction skill of JJAS area-averaged rainfall over Indian landmass region is higher in the perfect model scenario than the CFSv2_rev hindcasts (Table 2). We have found higher prediction skill in the perfect model scenario because bias and influence of initial shock is minimal in comparison to CFSv2_rev hindcasts.

13



R. P. Shukla et al.

Table 2  Correlation between predicted JJAS area-averaged rainfall with February initial conditions and May initial conditions during 21 cases and corresponding years in 50-years control simulation over Indian landmass region and same calculation has performed for real CFSv2_rev reforecasts for all cases (ensemble mean of 4, 8 and 12 members) during period 1979–2008

Feb IC May IC

Predictability experiments (perfect model scenario) Ensemble mean of five members

Real CFSv2_rev (1979–2008) reforecasts

Ensemble mean of four members

Ensemble Ensemble mean of eight mean of twelve members members

0.54 0.73

0.44 0.34

0.49 0.40

0.49 0.39

5 Conclusions The present study evaluates the seasonal (JJAS; June to September) predictability of the Indian summer monsoon (ISM) rainfall using the NCEP CFSv2. From a 50-year CFSv2 free simulation, we have chosen 21 cases that include typical wet, dry and normal ISM monsoon years as target years and conducted a set of predictability experiments with initial states in February and May. For each prediction, we have generated a five-member ensemble with perturbed atmospheric initial states and all predictions are integrated to the end of September. It is demonstrated that, based on the correlation coefficient and RMSE of JJAS seasonal mean rainfall, zonal wind at 850 hPa and sea surface temperature anomalies, the prediction skill is higher in predictability runs initialized with May IC than Feb IC over the ISM region. Based on an area-averaged summer rainfall index over the Indian landmass and extended area ISM rainfall indices, it is found that May IC predictability runs have higher correlation and lower RMSE than those with Feb IC, which implies that the initial states with the shorter lead (May) generally have the higher skill for predicting the summer mean rainfall, U850 and SST over the ISM region. These predictability experiments are employed to understand the finding reported by some recent studies (Chattopadhyay et al. 2015; Saha et al. 2016) that, initialized from the realistic ocean, land and atmospheric states, the NCEP CFSv2 seasonal hindcasts generally show higher skill in predicting the ISM rainfall anomalies from February initial states than from May initial states. Based on the findings from the predictability experiments, we explored the hypothesis that stronger atmospheric initial imbalance (or initial shock) occurs in the NCEP CFSv2 hindcasts with May IC, which not only affects the atmospheric circulation in May but also has longer-term influences though the impact on the ocean. It is found that SLHF is huge in May

13

over the Arabian Sea in hindcasts initialized with May IC and 10 m-wind is also stronger over this region. The model also produces excessive rainfall in May over the Arabian Sea in CFSv2 hindcasts with May IC. Due to excessive SLHF and stronger surface 10  m-winds in May over the Arabian Sea, the CFSv2 hindcasts with May IC tend to produce lower SST in the Arabian Sea in June, mainly near the Somalia–Oman upwelling region. The impact from the initial shock in hindcasts of SLHF (rainfall) in May is 17.3 W/ m2 (2.1  mm/day) for CFSv2 hindcasts with May IC. On other hand, we have not found any impact from the initial shock over the Arabian Sea in these variables with May IC in the predictability runs. We have shown that a fundamental difference between the predictability runs and the NCEP CFSv2 seasonal retrospective forecasts is that the latter are strongly influenced by the model climate drift, including the initial shock (Zhang 2011a, b), while those influences are practically absent in the former. Therefore, this analysis supports the hypothesis that the stronger influence of the initial shock in the NCEP CFSv2 hindcasts with May initial states over the Arabian Sea is the main reason for its reduced prediction skill for ISM rainfall anomalies, in comparison to those with February initial states. The initial shock in CFSv2 hindcasts is not restricted to the Indian summer monsoon region, but is a more general phenomenon. The manifestation of initial shock in CFSv2 hindcasts may vary from place to place and from season to season. It may also be most significant in different variables due to different initial imbalance of the coupled system in different locations. The global behavior of initial shock in CFSv2 hindcasts will be discussed in a separate paper in preparation. Acknowledgements  Funding for this research work was provided by grants from the National Science Foundation (1338427), the National Oceanic and Atmospheric Administration (NA14OAR4310160), and the National Aeronautics and Space Administration (NNX14AM19G). This research is also supported by a grant from the National Monsoon Mission, Ministry of Earth Sciences, Government of India. Computing resources provided by the Extreme Science and Engineering Discovery Environment (XSEDE) division are gratefully acknowledged. The authors are grateful to three anonymous reviewers for their constructive comments and suggestions, which improved the quality of the manuscript significantly Compliance with ethical standards  Conflict of interest  The authors declare that they have no conflict of interest.

References Achuthavarier D, Krishnamurthy V, Kirtman BP, Huang B (2012) Role of Indian Ocean in the ENSO-Indian summer monsoon

Predictability and prediction of Indian summer monsoon by CFSv2: implication of the initial… teleconnection in the NCEP climate forecast system. J Clim 25:2490–2508 Adler RF, Huffman GJ, Chang A, Ferraro R, Xie PP, Janowiak J, Rudolf B, Schneider U, Curtis S, Bolvin D, Gruber A, Susskind J, Arkin P, Nelkin E (2003) The version-2 global precipitation climatology project (GPCP) monthly precipitation analysis (1979–present). J Hydrometeorol 4:1147–1167 Balmaseda M, Vidard A, Anderson D (2008) The ECMWF System 3 ocean analysis system. Mon Weather Rev 136:3018–3034. doi:1 0.1175/2008MWR2433.1 Behringer DW (2005) The global ocean data assimilation system (GODAS) at NCEP, paper presented at 11th Symposium on Integrated Observing and Assimilation Systems for the Atmosphere, Oceans, and Land Surface (IOAS-AOLS), Am. Meteorol. Soc., San Antonio, Tex Charney JG, Shukla J (1981) Predictability of monsoons. In: Lighthill J, Pearce RP (eds) Monsoon dynamics. Cambridge University Press, Cambridge, pp 99–109 Chattopadhyay R, Rao SA, Sabeerali CT, George G, Rao DN, Dhakate A, Salunke K (2015) Large-scale teleconnection patterns of Indian summer monsoon as revealed by CFSv2 retrospective seasonal forecast runs. Int J Climatol. doi:10.1002/ joc.4556 Chaudhari HS, Pokhrel S, Saha SK, Dhakate A, Yadav RK, Salunke K, Mahapatra S, Sabeerali CT, Rao SA (2013) Model biases in long coupled runs of NCEP CFS in the context of Indian summer monsoon. Int J Climatol 33:1057–1069 DelSole T, Shukla J (2010) Model fidelity versus skill in seasonal forecasting. J Clim 23(18):4794–4806 Findlater J (1969) A major low-level air current near the Indian Ocean during the northern summer. Quart J R Meteorol Soc 95:362–380 Gadgil S, Gadgil S (2006) The Indian monsoon, GDP and agriculture. Econ Political Wkly 41(47):4887–4895 Huang B, Kinter JL III (2002) Interannual variability in the tropical Indian Ocean. J Geophys Res 107(C11):3199. doi:10.1029/20 01JC001278 Huang B, Shukla J (2007a) Mechanisms for the interannual variability in the tropical Indian Ocean. Part I: the role of remote forcing from the tropical pacific. J Clim 20:2917–2936 Huang B, Shukla J (2007b) Mechanisms for the interannual variability in the tropical Indian Ocean. Part II: regional processes. J Clim 20:2937–2960 Huang B, Zhu J, Marx L, Wu X, Kumar A, Z.Z. Hu, Balmaseda MA, Zhang S, Lu J, E.K. Schneider EK, Kinter JL (2015) Climate drift of AMOC, North Atlantic salinity and arctic sea ice in CFSv2 decadal predictions. Clim Dyn 44:559–583. doi:10.1007/s00382-014-2395-y Izumo T, Montégut CB, Luo JJ, Behera SK, Masson S, Yamagata T (2008) The role of the western Arabian Sea upwelling in Indian monsoon rainfall variability. J Clim 21(21):5603–5623 Jiang X, Song Yang, Yueqing Li, Arun Kumar, Xiangwen Liu, Zhiyan Zuo, Bhaskar Jha (2013) Seasonal-to-Interannual prediction of the Asian summer monsoon in the NCEP climate forecast system version 2. J Clim 26:3708–3727 Joshi UR, Rajeevan M (2006) Trends in precipitation extremes over India. National Climate Centre, India Meteorological Department. Kalnay E et al (1996) The NCEP/NCAR 40-Year reanalysis project. Bull Am Meteorol Soc 77:437–471 Levine RC, Turner AG (2012) Dependence of Indian monsoon rainfall on moisture fluxes across the Arabian Sea and the impact of coupled model sea surface temperature biases. Clim Dyn 38:2167–2190. doi:10.1007/s00382-011-1096-z Levine RC, Turner AG, Marathayil D, Martin GM (2013) The role of northern Arabian Sea surface temperature biases in CMIP5

model simulations and future projections of Indian summer monsoon rainfall. Clim Dyn 41:155–172. doi:10.1007/ s00382-012-1656-x Lorenz EN (1963) Deterministic non-periodic flows. J Atmos Sci 20:130–141 Mulholland DP, Laloyaux P, Haines K, Balmaseda MA (2015) Origin and impact of initialization shocks in coupled atmosphere– ocean forecasts*. Mon Weather Rev 143(11):4631–4644 Parthasarathy B, Munot AA, Kothawale DR (1988) Regression model for estimation of Indian food grain production from Indian summer rainfall. Agric For Meteorol 42:167–182 Pokhrel S, Saha SK, Dhakate A, Rahman H, Chaudhari HS, Salunke K, Hazra A, Sujith K and Sikka, DR (2016) Seasonal prediction of Indian summer monsoon rainfall in NCEP CFSv2: forecast and predictability error. Clim Dyn 46(7–8):2305–2326. Rahmstorf S (1995) Climate drift in an ocean model coupled to a simple, perfectly matched atmosphere. Clim Dyn 11(8):447–458 Rajeevan M, Guhathakurta P, Thapliyal V (2000) New models for long range forecasts of summer monsoon rainfall over Northwest and Peninsular India. Meteorol Atmos Phys 73:211–225 Rajeevan M, Pa DS, Dikshit SK, and Kelkar RR (2004) IMD’s new operational models for long range forecast of south-west monsoon rainfall over India and their verification for. Curr Sci 86:422–431 Rao KG, Goswami BN (1988) Interannual variations of sea surface temperature over the Arabian Sea and the Indian monsoon: a new perspective. Mon Wea Rev 116:558–568 Rasmusson EM, Carpenter TH (1983) The relationship between eastern equatorial Pacific sea surface temperature and rainfall over India and Sri Lanka. Mon Wea Rev 111:517–528 Saha K (1970) Zonal Anomaly of sea surface temperature in equatorial Indian Ocean and its possible effect upon monsoon circulation. Tellus 22:403–409 Saha S et  al (2010) The NCEP Climate Forecast System reanalysis. Bull Am Meteorol Soc 91:1015–1057. doi:10.1175/2010B AMS3001.1 Saha SK, Pokhrel S, Chaudhari HS, Dhakate A, Shewale S, Sabeerali CT, Salunke K, Hazra A, Mahapatra S, Rao AS (2014a) Improved simulation of Indian summer monsoon in latest NCEP climate forecast system free run. Int J Climatol 34:1628–1641 Saha S, Shrinivas Moorthi, Wu X, Wang J, Nadiga S, Patrick Tripp, David Behringer, Yu-Tai Hou, Hui-ya Chuang, Iredell M, Michael Ek, Jesse Meng, Yang R, Malaquías Mendez P, Huug van den Dool, Qin Zhang, Wang W, Mingyue Chen, Emily B (2014b) The NCEP climate forecast system version 2. J Clim 27:2185–2208 Saha SK, S Pokhrel, Salunke K, Dhakate A, Chaudhari HS, Rahaman H, Sujith K, Hazra A, Sikka DR (2016) Potential predictability of Indian summer monsoon rainfall in NCEP CFSv2. J Adv Model Earth Syst 8:96–120. doi:10.1002/2015MS000542. Saji NH, Goswami BN, Vinayachandran PN, Yamagata T (1999) A dipole mode in the tropical Indian Ocean. Nature 401:360–363 Sausen R, Barthel K, Hasselmann K (1988) Coupled ocean-atmosphere models with flux correction. Clim Dyn 2(3):145–163 Shukla J (1975) Effect of Arabian sea-surface temperature anomaly on Indian summer monsoon: a numerical experiment with the GFDL mode. J Atmos Sci 32:503–511 Shukla J (1981) Dynamical predictability of monthly means. J Atmos Sci 38:2547–2572 Shukla J (1987) Interannual variability of monsoons. In: Fein JS, Stephens PI (eds) Monsoons. Wiley, New York, pp 632 Shukla RP, Bohua Huang (2015a) Interannual variability of the Indian summer monsoon associated with the air-sea feedback in the northern Indian Ocean. Clim Dyn. doi:10.1007/ s00382-015-2687-x

13

Shukla RP, Bohua Huang (2015b) Mean state and interannual variability of the Indian summer monsoon simulation by NCEP CFSv2. Clim Dyn. doi:10.1007/s00382-015-2808-6 Shukla RP, Kinter JL (2014) Simulation of the Asian Monsoon using a regionally-coupled global climate mode. Clim Dyn. doi:10.1007/s00382-014-2188-3 Shukla J, Misra BM (1977) Relationships between sea surface temperature and wind speed over the central Arabian Sea, and the monsoon rainfall over India. Mon Wea Rev 105:998–1002 Shukla J, Paolin DA (1983) The Southern oscillation and long range forecasting of the summer monsoon rainfall over India. Mon Wea Rev 11:1830–1837 Shukla RP, Tripathi KC, Pandey AC and Das IML (2011) Prediction of Indian Summer monsoon rainfall using Niño indices: a neural network approach. Atmos Res 102(1) 99–109. doi:10.1016/j. atmosres.2011.06.013 Slingo J, Palmer T (2011) Uncertainty in weather and climate prediction. Philos Trans R Soc Lond A Math Phys Eng Sci 369(1956):4751–4767 Smith MT, Reynolds WR, Peterson TC, Lawrimore J (2008) Improvements to NOAA’s historical merged land–ocean surface temperature analysis (1880–2006). J Clim 21:2283–2296 Walker GT (1923) Correlation in seasonal variations of weather, VIII. A preliminary study of world weather. Mem India Meteorol Dep 24(4):75–131 Wang B, Ding Q, Fu X, Kang I-S, Shukla J, Doblas-Reyes F (2005) Fundamental challenge in simulation and prediction of summer

13

R. P. Shukla et al. monsoon rainfall. Geophys Res Lett 32:L15711. doi:10.1029/20 05GL022734 Webster PJ and Yang S (1992) Monsoon and ENSO: selectively interactive systems. Quart J R Meteorol Soc 118:877–926 Webster PJ, Magana VO, Palmer TN, Shukla J, Tomas RA, Yanai M, Yasunari T (1998) Monsoons: processes, predictability and the prospectsfor prediction. J Geophys Res 103(14 451–14):510 Wu R, Kirtman BP (2004) Impacts of the Indian Ocean on the Indian summer monsoon-ENSO relationship. J Clim 17:3037–3054 Wu R, Kirtman BP (2005) Roles of Indian and Pacific Ocean air– sea coupling in tropical atmospheric variability. Clim Dyn 25:155–170 Xie S-P, Annamalai H, Schott F, McCreary JP Jr (2002) Structure and mechanism of South Indian Ocean climate variability. J Climate 15:864–878 Zhang S (2011a) Impact of observation-optimized model parameters on decadal predictions: Simulation with a simple pycnocline prediction model. Geophys Res Lett 38:L02702. doi:10.1029/201 0GL046133 Zhang S (2011b) A study of impacts of coupled model initial shocks and state-parameter optimization on climate predictions using a simple pycnocline prediction model. J Clim 24(23):6210–6226 Zhu J, Shukla J (2013) The role of air–sea coupling in seasonal prediction of Asia-Pacific summer monsoon rainfall. J Clim 26:5689–5697